Convention: every experiment module states its prior-work search

One line, in the module docstring, before anything else:

Prior work: [[memo-name]] — <what it already established>

or, when nothing was found:

Prior work: none found (docs_search "<query>", source_type=memory)

For changed or newly added modules, run the repository check with the affected paths:

python -m tools.prior_art check-experiments benchmarks/my_probe.py

Why this exists rather than a rule someone remembers

On 2026-07-28 an abstention investigation was run that had already been run on 2026-07-24, and reached the same conclusion — cosine AUC 0.78 against the earlier 0.753, entailment 0.59 against 0.648. Cost: roughly four hours and several hours of VPS CPU, including 777 queries through a cross-encoder, to re-derive something already written down.

The rule to search memory first already existed in CLAUDE.md. It was skipped, and nothing about the skipping was visible — a reader of the resulting probe could not tell "searched, found nothing" from "never searched". That is the actual defect: not the omission, but that the omission left no trace.

A rule whose compliance produces an artifact gets followed, because its absence is conspicuous. The pre-registration convention works for exactly this reason — a prediction committed before a run is checkable afterwards. This is the same mechanism applied to prior work.

Scope

Any new module under benchmarks/ or results/ that measures something. A PreToolUse hook in .claude/settings.json prompts at the moment of creation; this file is what the prompt is asking you to satisfy.

Not required for: edits to an existing probe (the hypothesis was already declared), harness plumbing, or scripts that only reshape data already measured.

What a reviewer checks

Open the file. If the first paragraph does not say what was searched and what was found, the search cannot be assumed to have happened — treat the experiment as potentially duplicating work, and search before spending on it.